Think about the last time you figured something out. Maybe you noticed the school canteen runs out of samosas on Fridays, so you started going earlier. You observed something. You learned from it. You reasoned about what that meant for next time. You made a decision. That is what intelligence looks like in four steps: observe, learn, reason, decide. People do these four things every day — and so do some machines, but in a very different way. The columns are the same; it is the rows beneath them, the path a person takes and the path a machine takes, that start to look different. That is exactly where this chapter begins.
AI, artificial intelligence, is a system that finds patterns in data and uses them to answer questions. The key word is LEARNS — more examples mean better answers. People learn from experience, from being told things, and from working things out in their heads. A machine that uses AI learns from data instead — large collections of examples that have been stored so the machine can study them. That is the core idea. Hold it in mind as you read on.
Not every machine that solves a problem is an AI. This is one of the most important ideas in the whole chapter. Here is the difference in one sentence. Some machines follow rules that a person wrote in advance. Other machines improve by studying examples. The first kind is called RULE-BASED, or automation. The second kind is LEARNING-BASED — the AI kind.
Nila lays six cards on the table. Each card describes a surface that gives an answer. She sorts them without saying a word. A helpdesk lookup, a library catalogue, a transit kiosk, a bell schedule, and a calculator all go in Pile A. Pile A is rule-based. "Give it the same input twice," Nila says, "and you get the same output twice." A playlist app goes alone in Pile B. It notices you skipped songs after school and starts suggesting something different. Pile B is learning-based. "The surface improves by studying examples," Nila says.
Dhruv asks a question. What would happen if the transit kiosk met a bus added to the route last week? That bus is not yet in its schedule. "It would give you nothing, or the wrong time," Nila says. "The rule doesn't know about the new bus. No person has updated it yet." A rule-based system is only as good as its last update.
But that does not make it worse. A calculator is rule-based on purpose. You need 47 + 13 to always equal 60, every time. Now imagine a playlist app that only followed fixed rules. It would give everyone the same songs. It would have no way to notice that your taste shifted after a long day. The right kind of system depends on the job.
First, you break a problem into smaller pieces. This happens before you choose between rule-based and learning-based. That first step has a name: DECOMPOSITION. Decomposition means splitting one big problem into smaller problems. You solve those smaller problems one at a time. Once a problem is broken down, you plan the steps for solving each piece. Designing that plan is called ALGORITHM DESIGN. The plan itself, the ordered list of steps, is an ALGORITHM. The same algorithm gives the same result on the same input.
Nila writes her canteen question at the top of a fresh page. Which days does the canteen run out of samosas? Then she breaks it down. Pick a week to watch. Write down what is sold each day. See which day runs out first. That breakdown is decomposition. Her list of steps is an algorithm. Choosing the right steps in the right order is algorithm design. Both rule-based and learning-based systems rely on algorithms. A rule-based system follows an algorithm a person wrote. A learning-based system uses an algorithm that helps it learn from examples.
A learning-based system studies examples. But how? And what kind of examples? There are three main ways. Each one is called a learning paradigm. On Card A, a student sends ten leaf photos to an AI. She labels each one "healthy" or "spotted." The AI studies the photos and labels together. Afterward, it can look at a new photo and say which label fits. That is SUPERVISED learning — training on examples that already carry the right answer.
On Card B, a researcher sends five hundred news headlines to an AI. There are no labels at all. The AI notices which headlines share words and topics. It sorts them into clusters on its own. That is UNSUPERVISED learning — finding structure with no answers attached.
On Card C, a game-playing AI tries a move and scores ten points. It tries another move and scores two. Over thousands of tries, it learns to prefer the moves that score higher. That is REINFORCEMENT learning — learning by doing, from a score after each try.
Dhruv asks which paradigm depends on labels the most. "Supervised," Nila says. "It needs labelled examples. Take them away, and it has nothing to connect inputs to answers." Unsupervised still works without labels — it just needs the data itself, unlabelled, to look for groupings in. Reinforcement needs no examples prepared in advance at all. It makes its own experience by trying, as long as there is something to try actions in. "Supervised learning depends on labels more than the others do," Dhruv writes on a sticky note.
Two more words belong on those mini-cards now. A MODEL is what an AI system builds when it studies examples — not a physical thing, but a set of patterns the system has found, stored in a way that lets it answer new questions. TRAIN, or TRAINING, is the process of building that model: when you give a supervised system its labelled examples, the system is training; once training is done, the model is ready to answer questions it has never seen before. In higher classes you will see how a trained model gets checked, by trying it on three separate sets of examples called train, validation, and test.
AI is not one single thing that does everything. It is used in different domains. A domain is an area of work. In each domain, AI has been built and trained to do a specific kind of job.
Here is a first look at three domains. Computer Vision works with images and video. One example is a system that spots crop disease in a photo before a farmer would notice. Natural Language Processing works with words, written or spoken. One example is a system that translates a message from one language to another. Data Science studies large collections of data for useful patterns. One example is a system that finds rainfall patterns to predict dry seasons. You will meet all three again in later chapters. For now, remember the shape: images, words, patterns in data.
AI already shows up across healthcare, education, transport, and communication. It studies medical images so doctors notice what is easy to miss. It suggests practice problems where a student is struggling. It plans routes and spots delays before they happen. It translates languages and filters unwanted messages.
None of it is magic. Each system was trained on examples and built a model. Some parts need fixed answers — those are rule-based. Some parts need to handle something new — those are learning-based. That is what this chapter has been about: seeing the difference, and knowing which system does which job.